Data protection & recovery
Geo-replication
Geo-replication is the copying of data between storage systems or sites that are geographically separated, so that a complete copy exists in more than one location. It protects against the loss of an entire site and can place data closer to the users or applications that read it.
How does geo-replication work?
A replication relationship is defined between a source and one or more destinations, usually per bucket or per data set. As objects are written, changed or deleted at the source, the same changes are applied at the destination. Replication can run one way, giving a primary site and a standby, or both ways, giving two active sites that each accept writes.
What is the difference between synchronous and asynchronous?
The mode determines when a write is acknowledged and therefore how much data a site loss can cost.
| Aspect | Synchronous | Asynchronous |
|---|---|---|
| Write acknowledged | After all sites have the data | After the local site has it |
| Data lost if a site fails | None | Whatever had not yet replicated |
| Write latency | Includes the inter-site round trip | Local only |
| Distance supported | Limited by latency, typically metro | Any distance |
| Effect of a link outage | Writes stall or the cluster degrades | Writes continue, a backlog builds |
What is replication lag?
In asynchronous replication, lag is the delay between a write at the source and its arrival at the destination. It defines the recovery point for a site failure: everything written during the lag is at risk. Lag grows when the write rate exceeds what the link can carry, during bursts or after an outage while the backlog drains. It should be measured and alerted on as an age, such as the oldest object not yet replicated, because a healthy link can still be minutes or hours behind. Recovery point objective is the target lag is compared with.
Where does geo-replication fit, and where does it not?
Geo-replication provides site resilience. It is not a backup. A deletion, an overwrite or corruption at the source is replicated faithfully to the destination, so protection against those needs versioning, immutability or separate backups. Location rules also apply: replicas must stay within any territory the data is restricted to, as under data localization. The synchronous case is often built as a stretched cluster. Cross-border replication and site failover cover the legal and operational sides.
Frequently asked questions
Is geo-replication the same as backup?
No. Replication copies changes, including mistakes and deletions. Backup keeps earlier points in time.
How far apart should replicated sites be?
Far enough that one event cannot affect both, and within the distance the chosen mode tolerates. Synchronous replication is usually limited to metropolitan distances.
Does geo-replication double the storage cost?
Each full replica needs its own capacity. Some designs spread erasure-coded data across three or more sites to reduce the overhead.
Related terms
- Data replication: the general technique.
- Stretched cluster: one synchronous cluster across sites.
- Multi-region storage: data placed across regions.
- Recovery point objective (RPO): the target replication lag is measured against.
- Data localization: location limits on where replicas may be.